Skip to main content
Semantic search helps you uncover markets related to a topic, even when titles differ. Tremor maintains embeddings for Polymarket and Kalshi using Qwen 8B vectors served through DeepInfra. 
curl -X POST https://tremor.sh/api/markets/semantic-search \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "election markets with high volume",
    "platform": "polymarket",
    "limit": 5,
    "only_active": true
  }'

​
Request fields

  • query: Natural-language description of what you need (required)
  • platform: Optional filter (polymarket or kalshi)
  • limit: Defaults to 20; capped at 50
  • min_volume: Screen out thin markets
  • only_active: Return only markets still trading

​
Response Primer

Each result includes similarity scores plus live market metrics: 
{
  "result_type": "market",
  "platform": "polymarket",
  "title": "Will the EU pass an AI Act by 2025?",
  "category": "Politics",
  "ticker": "ai-act-2025",
  "similarity": 0.81,
  "volume": 123456.78,
  "yes_bid": 0.62,
  "yes_ask": 0.64,
  "status": "active"
}
Use the similarity score (1.0 is a perfect match) to rank or threshold results.

​
Enrich with SQL

Once you have a shortlist, feed market identifiers into SQL queries for deeper analysis: 
SELECT
  m.title,
  m.yes_probability,
  m.volume_24hr,
  m.liquidity
FROM polymarket_events_latest m
WHERE m.id IN {{comma-separated IDs from semantic search}}
ORDER BY m.volume_24hr DESC;
Combine with the Query Recipes snippet to merge vector results and historical snapshots.

​
How It Works

  • Embeddings are refreshed continuously via the railway-sync workers.
  • Tremor stores vectors in ClickHouse tables (*_market_embeddings) for sub-200 ms search latency.
  • cosineDistance scores return similarity; Tremor converts to 1 - distance before responding.
See the API reference for full parameter and schema details.